Research And Framing

The design grounds itself in specific lines of recent research.

MemReranker established that generic rerankers fail in memory settings because semantically similar fragments often lack answer-bearing information. Qwen3-Reranker provides a practical small open-weight model that is CPU-deployable. Embedding-retrieval theory work showed structural limits on single-vector retrieval as corpora grow, which is why CoreTex combines many compact retrieval keys with relation routes, a codebook, and reranking. LoCoMo and MemoryAgentBench established temporal current-versus-stale retrieval as its own evaluation problem. The Experience Compression Spectrum frames memory, skills, and rules as different compression levels, and the 32 KB substrate is a protocol-level compression surface.

The consistent theme: retrieval quality is what makes memory better.